Facebook has revealed in a research paper that it has discovered a way to match faces from two different photographs with 97.25 percent accuracy. The research has aided in the development of a software project called DeepFace, which only slightly trails the
facial recognition measured average of a human by only 0.28 percent.
The increased accuracy comes at a rate nearly 25 percent higher compared to other software by a process of facial verification rather than recognition. Improvement in the rate of verification is independent of changes in angle of the face towards a camera or differences in lighting. "You normally don't see that sort of improvement," says Yaniv Taigman, one of the researchers on Facebook's AI team.
Analysis from
MIT on the claims made in the
research paper yields information on the changing algorithms in facial recognition through the use of deep learning. The process of deep learning, in which artificial intelligence learns patterns in data through a network of simulated neurons, creates a numerical description of a photographed face which has been reoriented to face forward with use of a 3D model of an averaged forward-looking face as a guide. If enough matches are generated after the photos being compared are found, the software indicates they must be the same.
Facebook was able to train the layers of neurons in the deep learning system by using a sample of the photos stored by the company. Approximately four million photos of faces from around 4,000 users were accessed in the process.
The software was developed by a team of four, including Taigman, Ming Yang and Marc'Aurelio Ranzato of Facebook's internal research department and Lior Wolf, a professor at Tel Aviv University. The software is currently only a research project, and there are no plans for it to be deployed for widespread use at this time. Findings of the paper will be presented at the IEEE conference on Computer Vision and Pattern Recognition in June.